Records

Finding the best single opposing setup against the pool and the prior

Registration 9fbb92bf-2969-4683-855f-6d0721346e94 (study setups/selection-by-self-play), run 2026-09-19 on AMD Ryzen AI Max (Strix Halo), 32 threads, 125 GB unified memory, Radeon 8060S via ROCm, Linux 6.18. Server commit 0eb994c, research commit 19c62a6. Both runs shared the machine with other agents' jobs, which changes wall time only; the procedure is deterministic under its seed and the output files never carry the wall clock, so a rerun with the same arguments is byte-identical (tested in server/crates/arena/tests/setups.rs).

Commands:

cd server && target/release/setups exploit --policy 'expectimax:{"samples":16,"depth":2}' --against ../research/content/assets/setup-pool.json --n 1024 --games 8 --seed 20260920 --rules training --threads 8 --out ../research/runs/setup-selection/exploiter-pool.json
cd server && target/release/setups exploit --policy 'expectimax:{"samples":16,"depth":2}' --against structured --n 1024 --games 8 --seed 20260921 --rules training --threads 8 --out ../research/runs/setup-selection/exploiter-prior.json

The exploit procedure is the selection procedure run against a fixed source down to one setup: 1,024 candidate draws of the prior, 8 paired games per round per candidate against the source, successive halving, mutants from the second round, ten rounds and 16,368 games per run. Both runs play expectimax over sixteen worlds at depth 2 in both seats under the training rules.

Both winners are mutants of earlier candidates. The procedure scores are cumulative in-procedure measurements, not the exploitability proxies; the proxies are the two 400-paired-game evaluation cells (pool, prior) registered beside this one.

Provenance

Archived output: runs/setup-selection/exploiter-pool.json (SHA-256 461a099c27c8951ae3b2edff6337074fd8f8dfcbdbe4329590376753a87a3a29) with log runs/setup-selection/exploit-pool.log (SHA-256 0d56dce1cf03766207221d8d23cad495c3e1eb43f1735958815aea6eaea3dc2c), and runs/setup-selection/exploiter-prior.json (SHA-256 3eb714a43f83836db0aa729809501025a3ad41a13f437bbec797629302e60255) with log runs/setup-selection/exploit-prior.log (SHA-256 ae1a7675f38cb4a95bc1acbf6adf98d42e2283091bd20d9ab25ccffa694d2740). Games planned 32,736, attempted 32,736, completed 32,736, interrupted 0. Each procedure completed its ten rounds and wrote one legal setup with its cumulative effective score, its 95% Wilson interval and its game count. No strength claim is decided here; the evaluation cells decide it.